Research On Radar Signal Modulated Recognition Method Based On Deep Learning

Radar signal modulation recognition is an important part of electronic warfare reconnaissance.By identifying the type of radar signal modulation,important information such as the model,purpose,and threat level of the transmitting signal equipment can be obtained,which is of great significance for electronic warfare.With the increasingly complex electronic warfare environment and the rapid development of radar technology,the widely used deep learning based radar signal modulation recognition method mainly has two problems.Firstly,deep learning based recognition methods require a sufficient number of samples to serve as a training set in order to converge the network model after iterative training and achieve good recognition results.However,in the current complex environment,it is difficult to collect certain or even certain types of radar signal samples,which can cause insufficient samples and seriously reduce recognition performance;Secondly,for new unknown radar signals outside the training set,deep learning based recognition methods cannot effectively distinguish between unknown signals and known signals,and even affect the recognition accuracy of known signals due to the addition of unknown signals.In response to the above issues,this article first proposes Weighted Adaptive Blur Deep Convolution Generative Adversarial Networks(WAB-DCGAN)to solve the problem of small sample recognition.Secondly,a recognition method based on deep metric learning is proposed to solve the problem of unknown signal recognition outside the training set.The main research content is as follows:1.The paper briefly describes the commonly used radar signals,provides mathematical models,and analyzes the time-frequency characteristics of different signals through simulation.At the same time,the advantages and disadvantages of the three time-frequency analysis methods used in this article are summarized.Finally,a brief overview of the structure and principles of Convolutional Neural Networks(CNN)is provided,and the training process of CNN is provided.2.Propose a small sample recognition method based on WAB-DCGAN to solve the problem of lacking radar signal samples in complex environments.Because the cross entropy loss function is prone to gradient dispersion,leading to a decline in the learning rate,first replace the cross entropy loss function in Deep Convolutional Generative Adversarial Networks(DCGAN)with the weighted cross entropy loss function.The weighted cross entropy loss function can,on the one hand,rely on the constant weight coefficient to artificially significantly increase the contribution of the low number of categories to the loss function,on the other hand,inhibit the contribution of the high number of categories to the loss function,Therefore,we can balance the contributions of original samples and real samples to the loss function in the training phase,so that the training phase can be more stable;Secondly,adaptive blur is added before the original and generated samples reach the decision maker to improve the decision maker’s ability and indirectly improve the generator’s ability.By relying on the game between the generator and the decision maker,the generated samples of the generator are as close as possible to the real samples,in order to compensate for the problem of insufficient radar signal samples;Through simulation experiments,it can be seen that high recognition accuracy can be achieved whether it is adding a certain number of generated samples to a small number of fixed original samples,or adding a certain number of real samples to a small number of fixed generated samples;Under the same experimental conditions,the performance of WAB-DCGAN is superior to that of DCGAN.3.This paper proposes a deep metric learning based recognition method for new unknown radar signals that appear outside the training set in complex electromagnetic environments,combined with metric learning methods.First,Cosine Softmax is used to replace the traditional Softmax layer.This structure can bring better intra class aggregation and inter class separation.It can make better use of features in the training phase to make the same classes gather more closely and different classes are farther apart.Second,three time-frequency analysis images are used as inputs for simultaneous training,at the training stage,cross entropy loss function(using Cosine Softmax),magnet loss(using Softmax)and triplet loss(using Softmax)are used for comparison.The neural network learns the input and maps it to a specific embedded space.Through the above three training strategies,the features of similar radar signals are closely clustered,and the features of different radar signals are sufficiently dispersed.Finally,simulation experiments show that,on the premise of high recognition accuracy for known radar signals,it has good recognition accuracy for unknown radar signals.